app.py

import os
import json
import gradio as gr
from mistralai.client import Mistral


api_key = os.environ["MISTRAL_API_KEY"]
client = Mistral(api_key=api_key)
model_name = "mistral-large-latest"


def generate_MCQquestion(text):
    prompt = f"""
    You are a strict Arabic exam generator.
    You MUST use ONLY the information explicitly written in the Context.
    DO NOT use any external knowledge.

    Context:
    {text}

    Generate MCQs in Arabic based on the length of the Context.

    Question count rules:
    - Very short context (< 100 words): 3 questions.
    - Short context (100–300 words): 3–5 questions.
    - Medium context (300–700 words): 5–7 questions.
    - Long context (> 700 words): 7–10 questions.

    STRICT RULES FOR OPTIONS & ANSWERS:
    1. **Randomize Answer Position:** The correct answer index (0-3) MUST be distributed randomly across the questions. Do NOT always place the correct answer at index 0.
    2. **Length Consistency:** All 4 options for a single question should have roughly the same length. Do NOT make the correct answer significantly longer or more detailed than the distractors.
    3. **Plausible Distractors:** Incorrect options must be highly plausible and derived from the context's vocabulary to challenge the student.
    4. **Source Integrity:** Every correct answer and its "answer_source" MUST exist explicitly in the text.

    STRICT QUALITY RULES:
    - Quality over quantity.
    - No outside knowledge or general info.
    - If the context doesn't support the required number of questions, stop at what is factual.

    Return ONLY valid JSON:
    {{
      "mcqs": [
        {{
          "question": "string",
          "options": ["opt1", "opt2", "opt3", "opt4"],
          "answer_index": integer (0 to 3),
          "answer_source": "exact sentence from context"
        }}
      ]
    }}
    """

    try:

        response = client.chat.complete(
            model=model_name,
            messages=[
                {"role": "user", "content": prompt}
            ],
            response_format={"type": "json_object"},
            temperature=0.2
        )

        content = response.choices[0].message.content
        data = json.loads(content)

        # handle both dict and list formats
        if isinstance(data, dict):
            mcqs = data.get("mcqs", [])
        elif isinstance(data, list):
            mcqs = data
        else:
            return json.dumps({"error": "Invalid response format"}, ensure_ascii=False)

        # enforce limits
        if len(mcqs) < 3:
            return json.dumps({"error": "Model generated fewer than 3 questions"}, ensure_ascii=False)

        if len(mcqs) > 10:
            mcqs = mcqs[:10]

        questions = []
        answers = []

        for i, item in enumerate(mcqs, start=1):

            # safety checks
            if (
                "question" not in item or
                "options" not in item or
                "answer_index" not in item
            ):
                continue

            if not isinstance(item["options"], list) or len(item["options"]) != 4:
                continue

            if item["answer_index"] < 0 or item["answer_index"] > 3:
                continue

            questions.append({
                "id": i,
                "question": item["question"],
                "options": item["options"]
            })

            answers.append({
                "question_id": i,
                "correct_answer_index": item["answer_index"] + 1
            })

        final_output = {
            "questions": questions,
            "answers": answers
        }

        return json.dumps(final_output, ensure_ascii=False, indent=2)

    except Exception as e:
        return json.dumps({"error": str(e)}, ensure_ascii=False, indent=2)


# UI
demo = gr.Interface(
    fn=generate_MCQquestion,
    inputs=gr.Textbox(lines=10, label="أدخل النص"),
    outputs=gr.Code(language="json", label="JSON Output"),
    title="Arabic MCQ Generator API Ready",
)

demo.launch()